HECATE: New complexity metrics for LLM-integrated applications

HECATE, introduced by Zihao Xu, Yuekang Li, Gelei Deng, Yi Liu and Zhenchang Xing, assesses complexity across both prompt and code layers of LLM-integrated applications. Submitted on 2 Jul 2026, the paper presents a tool and methodology that generated 52 candidate metrics and retained ten that predict maintenance activity.

What is HECATE and how does it measure complexity?

HECATE is a tool that treats prompts as first-class specifications and measures complexity in both prompts and code. The authors formalize prompts with a Hoare-logic-inspired approach called "Prompt-as-Specification" and derive metrics from 25 complexity dimensions found in published taxonomies, producing 52 candidate metrics for evaluation.

HECATE emphasizes structural measures rather than sheer volume. Seven of the ten surviving metrics are newly introduced by the authors and count structurally distinct elements such as LLM call sites, memory attributes, and prompt templates, a property the paper names structural breadth. The remaining three surviving metrics are conventional: RFC, Halstead N, and Halstead V, with RFC showing a breadth-oriented character while Halstead N and V survive primarily as a residual effect of size.

How was HECATE validated?

The authors validated HECATE across historical maintenance signals from open-source projects, testing metrics on a concrete, repository-level dataset. They evaluated 118 components collected from 18 open-source repositories and used maintenance activity derived from version history as an empirical proxy for complexity.

Each candidate metric was assessed for significance against that maintenance proxy, and any metric that lost statistical significance once code size was accounted for was discarded. After this filtering, ten metrics remained significant. The authors performed a final validation on 20 components spanning six held-out repositories, and report that the two best-performing metrics continued to predict maintenance effort on those held-out components, supporting generalizability beyond the training set.

Why it matters

HECATE reframes complexity measurement for software that embeds natural language prompts by showing prompt-layer signals matter independently from code size. The analysis finds that prompt-layer metrics retain significance even when the strongest code-level metric is included as a covariate, establishing prompt complexity as a distinct dimension. That implies teams measuring or managing maintenance effort for LLM-integrated applications must look beyond code-level statistics and include prompt-structure measures.

The paper also supplies a practical shortlist of metrics: out of 52 candidates grounded in 25 dimensions, only ten passed the authors' significance filter, and seven of those are new, structurally focused metrics. RFC and two Halstead measures remain relevant but do not replace prompt-layer signals.

What to watch

Watch for broader adoption of prompt-layer metrics in empirical software engineering: the paper’s next confirmatory step would be replication across larger, diverse corpora of LLM-integrated projects. Also track whether toolchains and linters begin reporting structural-breadth measures such as counts of LLM call sites, memory attributes, and prompt templates.

Methodological specifics and key data points

• Submission date: 2 Jul 2026.
• Candidate metrics generated: 52, grounded in 25 published complexity dimensions.
• Components evaluated in primary study: 118 from 18 open-source repositories.
• Surviving metrics after significance and size controls: 10 total; 7 newly introduced, 3 conventional (RFC, Halstead N, Halstead V).
• Held-out validation: 20 components across six repositories, where the two best-performing metrics continued to predict maintenance effort.

The paper positions HECATE as the first tool designed to assess complexity in both prompt and code layers of LLM-integrated applications and introduces concrete, validated metrics teams can consider when analyzing maintenance risk.